一种适于非特定人语音识别的并行隐马尔可夫模型

An Appropriate Parallel HMM for Speaker-Independent Speech Recognition

为了适合非特定人语音识别,提出了一种由多条并行马尔可夫链组成的并行HMM(Parallel Hidden Markov Model,PHMM),从而融合了基于分类的语音识别中为各个类别建立的模板,提高了识别性能,各条链之间允许有交叉,使得融合的多模板之间存在状态共享,同时PHMM可以在训练过程中自动完成聚类,且测试语音的输出结果来自所有类别,无需聚类分析和类别判断,这些都减少了存储量和计算量,汉语非特定人孤立数字的识别实验表明,PHMM较之传统CHMM使识别性能及噪声鲁棒性都得到了改善。

In this paper Parallel Hidden Markov Model(PHMM)made up of several parallel Markov chains is proposed to fit in with speaker-independent speech recognition. The performance is improved because of the fusion of different models from classification based speech recognition. By sharing states of fused models, making classification automatically during training and getting result from all classifications, the amount of storage and operation can be decreased. The experiment for speaker-independent recognition of mandarin isolated digit shows that the PHMM improves the recognition performance and noise robustness.